3D dose prediction for Gamma Knife radiosurgery using deep learning and data modification.

Area of Science:

• Medical Physics
• Radiosurgery Technology
• Artificial Intelligence in Medicine

Background:

• Accurate 3D dose prediction is crucial for optimizing Gamma Knife (GK) radiosurgery outcomes.
• Existing methods may face challenges with the variability in target number, size, and shape.
• Machine learning offers potential for enhancing dose prediction accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a machine learning-based 3D dose prediction methodology for Gamma Knife (GK) radiosurgery.
• To ensure the methodology accommodates targets of any number, size, and shape.
• To compare the performance of GK-specific deep learning models against standard models.

Main Methods:

• A dataset of 322 GK treatment plans was modified, including contouring, cropping, and scaling of target areas to create 3D tensors.
• Generative Adversarial Network (GAN-GK) and 3D U-Net (U-Net-GK) models were trained on the modified dataset (272 patients).
• Model predictions for 50 out-of-sample patients were evaluated using gamma analysis (4%/2mm, 3%/3mm, 3%/1mm, 1%/1mm) and quality indices.

Main Results:

• GK-specific GAN and U-Net models achieved high average gamma passing rates (4%/2mm) of 84.9% and 83.1%, respectively, comparable to clinical plans.
• Baseline models trained on unmodified data showed significantly poorer prediction accuracy across all gamma criteria (p < 0.001).
• The quality indices (coverage, selectivity, conformity) of GK-specific predictions were similar to clinical dose distributions.

Conclusions:

• Deep learning models, when trained with GK-specific data modifications, can accurately predict 3D dose distributions for Gamma Knife Radiosurgery (GKRS) plans.
• This approach effectively handles variations in target characteristics, including number, size, and shape.
• Standard deep learning models applied to unmodified data are less effective for GKRS dose prediction.